yeast.wiki — a hosted, bilingual implementation of Karpathy's LLM-wiki idea

In a widely-shared gist, Andrej Karpathy sketched a simple but powerful idea. yeast.wiki is an implementation of it — made hosted, multi-tenant, bilingual, and own-your-data.

The idea

Instead of retrieval-augmented generation (RAG) — where a model fetches chunks and answers each query from scratch — the gist proposes an LLM that incrementally maintains a synthesized, interlinked Markdown wiki from your raw sources. As you feed it material, it integrates the new information into existing pages, writes new ones, and keeps the cross-links current. Knowledge compounds into a durable, navigable wiki rather than evaporating as one-off answers.

A wiki the model keeps writing and re-linking from your sources — not a chatbot you re-query, and not a pile of notes you hand-organize.

The idea resonated: it spawned many open-source re-implementations within weeks — mostly single-user, local, English-only CLIs and scripts.

What yeast.wiki adds

yeast.wiki takes that core loop — ingest → integrate → link → maintain — and builds the product around it:

• Hosted & multi-tenant. Nothing to run locally; your wiki is always on and reachable from any device.
• Bilingual EN / 中文, first-class. Pages are written in your primary language; translations are produced on demand and kept fresh — a gap in nearly every other implementation.
• Own your data. Plain Markdown you keep — synced to Obsidian over WebDAV and backed up to your own GitHub. No lock-in.
• More than one way in. Feed it and read it via the web, Telegram, and an in-page chat — and (the gist's actual thesis) build the wiki from documents, not just dictation.
• A resident agent ("Yeasty") that tends the wiki continuously, rather than a script you invoke.

The same thesis: a wiki from sources, not RAG

The gist's sharpest point is "a wiki built from your sources, instead of RAG." That's exactly where yeast.wiki invests: importing your documents (PDFs and more) and synthesizing them into linked, bilingual pages you own — so the knowledge accumulates and connects over time, and you can explore it as a map rather than re-asking a model.

Try it

A private, self-organizing, bilingual AI wiki you own. Invite-only.

Open your wiki